Wonder3D
github.com
- Free plan
- Windows
- Linux
- Web
- Self-hosted
What Wonder3D is
Wonder3D is a free project that reconstructs textured 3D meshes from a single-view image. It generates consistent multi-view normal maps and color images with cross-domain diffusion, then offers Instant-NSR and NeuS mesh-extraction workflows. The project reports reconstruction taking about 2–3 minutes and provides local Gradio demos for normal, color, and mesh generation. The current implementation supports six views and 256×256 generation, resizing inputs to that resolution. It warns that occluded objects may reconstruct worse when six views cannot cover the complete shape, and recommends front-facing images with fewer occlusions. Foreground preparation can use Clipdrop or rembg, while inference requires downloaded checkpoints including a Segment Anything Model checkpoint. The repository includes two training scripts and describes a two-stage training process. NeuS uses less GPU memory and favors smooth surfaces but takes longer and may produce less sharp textures. Wonder3D is MIT-licensed and free with acknowledgement. Documented platforms include Linux, Windows through a separate setup branch, web, and self-hosted use.
Who it suits
It suits creators working on 3D assets, Houdini, 3D printing, 2D designs, or animal imagery. Users comfortable with local model checkpoints, segmentation, and the documented six-view limits are the clearest fit.
Good to know, and what to watch
- Reconstructs textured meshes from one image
- Provides Instant-NSR and NeuS workflows
- Includes local Gradio demos
- Free under the MIT license
- Supports image-to-3D and AI texturing
- Generation is limited to six views
- Inputs are resized to 256×256
- Occluded objects can reconstruct worse
- Inference requires downloaded model checkpoints
The verdict
Wonder3D offers a free image-to-3D workflow with mesh extraction, local demos, and documented training resources. Its six-view, 256×256 setup and occlusion caveat are important constraints when preparing inputs.
What the maker publishes
- Contact and collaboration
- The project invites collaboration and lists [email protected] for contact. source
- Founded
- 2023 source
- Generation
- It generates consistent multi-view normal maps and corresponding color images using cross-domain diffusion. source
- Generation method
- It generates consistent multi-view normal maps and corresponding color images with a cross-domain diffusion model, then uses normal fusion for reconstruction. source
- Hardware
- The example pipeline moves to CUDA when available, and the repository provides a one-GPU inference configuration. source
- Input guidance
- The README says front-facing images with fewer occlusions generally produce better reconstructions. source
- Input preparation
- The README suggests using Clipdrop or rembg to segment the foreground object. source
- Integrations
- The documented workflow uses Hugging Face Diffusers and checkpoints, and describes using SAM and Clipdrop or rembg for foreground masks. source
- Intended uses
- The project page shows examples involving Houdini, 3D asset creation, 3D printing, 2D designs, and animal images. source
- License
- The repository states that Wonder3D is under the MIT license and free to use with acknowledgement. source
- Local demos
- The repository provides local Gradio demos for generating normals and colors and for mesh reconstruction. source
- Mesh extraction
- It provides Instant-NSR and NeuS-based mesh extraction workflows. source
- Model dependencies
- The inference instructions call for downloaded checkpoints and a Segment Anything Model checkpoint. source
- Multi-view attention
- The method uses multi-view cross-domain attention to exchange information across views and image modalities. source
- Notable limits
- The README says the implementation supports six views and 256×256 generation, and resizes input images to that resolution. source
- Output limits
- The current implementation supports six views and resizes input images to 256×256 for generation. source
- Platforms
- The README provides Linux setup, points Windows users to a separate branch, and links a Hugging Face demo. source
- Processing time
- The project says reconstruction takes about 2–3 minutes. source
- Purpose
- Wonder3D reconstructs textured 3D meshes from a single-view image. source
- Reconstruction caveat
- The README says occluded inputs can produce worse reconstructions because six views may not cover the complete object. source
- Reconstruction options
- The repository documents Instant-NSR and NeuS mesh extraction; it says NeuS uses less GPU memory and favors smooth surfaces but takes longer and may produce less sharp textures. source
- Reconstruction speed
- The project says reconstruction takes about 2–3 minutes. source
- Support and collaboration
- The README invites questions and collaboration by email at [email protected]. source
- Supported systems
- The README provides Linux setup instructions and points Windows users to a separate setup branch. source
Checked against the maker's pages on 7 Oct 2026.